Embedding-Based Counterfactuals: Using Deep Network Embeddings to Simulate and Test Hypothetical Social Interventions

Recent literature (Jing Ma, 2024; Brand et al., 2023) points to the potential of machine learning and embeddings for causal inference, but mainly for effect estimation or prediction. This idea takes it further: train deep network embedding models (e.g., node2vec, GNNs) that capture rich structural and attribute information from observed social networks; use these models to generate plausible counterfactual network states—e.g., simulating the removal of certain nodes (influencers), the rewiring of ties, or the injection of new cross-group links; run causal inference procedures (such as those in Tian & Rizoiu, 2025) on both real and counterfactual networks to estimate the potential impact of interventions (e.g., “What if we connected these two communities?” or “What if this influencer had not posted?”). This approach brings together cutting-edge representation learning with causal reasoning to enable “what-if” scenario analysis at scale, supporting both academic understanding and real-world policy design in online platforms, health, or political communication.

References:

1. Estimating Online Influence Needs Causal Modeling! Counterfactual Analysis of Social Media Engagement. Lin Tian, Marian-Andrei Rizoiu (2025). arXiv.org.
2. When Causal Inference Meets Graph Machine Learning. Jing Ma (2024). AAAI Conference on Artificial Intelligence.
3. Recent Developments in Causal Inference and Machine Learning. J. Brand, Xiaoping Zhou, Yu Xie (2023). The Annual review of sociology.